Joint Learning of Representations for Web-tables, Entities and Types using Graph Convolutional Network

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1197–1206 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1197 Joint Learning of Representations for Web-tables, Entities and Types using Graph Convolutional Network Aniket Pramanick TCS Research aniket.pramanick@tcs.com Indrajit Bhattacharya TCS Research b.indrajit@tcs.com Abstract Existing approaches for table annotation with entities and types either capture the structure of table using graphical models, or learn em- beddings of table entries without accounting for the complete syntactic structure. We pro- pose TabGCN, which uses Graph Convolu- tional Networks to capture the complete struc- ture of tables, knowledge graph and the train- ing annotations, and jointly learns embeddings for table elements as well as the entities and types. To account for knowledge incomplete- ness, TabGCN’s embeddings can be used to discover new entities and types. Using ex- periments on 5 benchmark datasets, we show that TabGCN significantly outperforms multi- ple state-of-the-art baselines for table annota- tion, while showing promising performance on downstream table-related applications. 1 Introduction Table data abounds in webpages and organizational documents. Annotation of table entries, such as columns, cells and rows, using available back- ground knowledge (e.g. Yago, DBPedia, Freebase, etc.), such as knowledge of entities and their types, helps in better understanding and semantic interpre- tation of such tabular data. The challenge, however, is that such web tables do not adhere to any stan- dard format, schema or convention (Limaye et al., 2010). Additionally, knowledge graphs are typ- ically incomplete - entities and types mentioned in tables may not always exist in the knowledge graph. Therefore, it becomes necessary to expand the knowledge graph with new entities (Zhang et al., 2020) and types for annotating tables. Initial research on table annotation (Limaye et al., 2010; Takeoka et al., 2019; Bhagavatula et al., 2015) used probabilistic graphical models to capture the complete row-column structure of tables and also the knowledge graph for collective annotation. More recent approaches using embed- dings (Gentile et al., 2017; Zhang and Balog, 2018; Zhang et al., 2019; Chen et al., 2019; Yin et al., 2020) only partly capture the syntactic structure of tables, and also ignore the structure of the knowl- edge graph. The problem of incompleteness of the knowledge representation (Zhang et al., 2020) is mostly not addressed. In this work, we propose the TabGCN model that uses a Graph Convolutional Network (GCN) (Kipf and Welling, 2017) to unify the complete syntactic structure of tables (rows, columns and cells) and that of the knowledge graph (entities and types) via available annotations. The embeddings of the table elements as well as knowledge graph entities and types are trained jointly and end-to-end. While GCNs have been used for learning embed- dings for many NLP tasks using the syntactic and semantic structure of natural language sentences (Marcheggiani and Titov, 2017; Vashishth et al., 2019), encoding tabular structure using GCNs has not been addressed before. The model and embed- dings thus trained are used to annotate new tables with known …

Content truncated to 500 words